On the Use of Artificial Neural Network Technique for Seismic Fragility Analysis of a Three-Dimensional Industrial Frame

Abstract

Abstract Fragility function that defines the probability of exceedance of a damage state given a ground motion intensity (IM) is an essential ingredient of modern approaches to seismic engineering as the performance-based earthquake engineering methodology. Epistemic as well as aleatory uncertainties associated with seismic loads and structural behavior are usually taken into account to analytically develop such curves. However, structural analyses are time-consuming, requiring generally a high computational effort. Moreover, the conditional probability of failure is usually computed by regression analysis assuming predefined probability functions, like the log-normal distribution, without prior information on the real probability distribution. To overcome these problems, the artificial neural network (ANN) technique is used for the development of structural seismic fragility curves considering record-to-record variability and structural parameter uncertainties. In this respect, the following aspects are addressed in this paper: (a) implementation of an efficient algorithm to select IMs as inputs for ANN, selecting the most relevant ones; (b) derivation of surrogate models by using the ANN technique, c) computation of fragility curves with Monte Carlo Simulations method and verification of the validity. These methods enable the implicit treatment of uncertainty in either or both of ground motion intensity and structural properties without making any prior assumption about the probability function. This methodology is then applied to estimate the probability of failure of a non-structural component (NSC), i.e., vertical tank, located on a typical three-dimensional industrial frame. First, an extensive sensitivity analysis on the ANN input parameters is performed (feature selection), identifying the type and number of seismic intensity measures (amplitude-based, frequency-based, and time-based IM). Then different surrogate models are derived investigating the number of hidden layers and parameters. A multiple stripe analysis is then performed on a nonlinear model of the structure, deriving the set of data for the ANN. Different training and test subsets are used to derive the surrogate model. Finally, a Monte Carlo simulation is performed to derive the fragility curves for the limit state considered. Finally, the risk assessment is obtained, evaluating the mean annual rate of failure of the NSC.